MFE-Former: Disentangling Emotion-Identity Dynamics via Self-Supervised Learning for Enhancing Speech-Driven...

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Acoustic features in speech are vital for depression detection.
• Existing methods struggle with emotional variability and speaker identity interference.
• Accurate detection of emotional changes is crucial for understanding depression.

Purpose of the Study:

• To develop a robust speech-based depression detection method addressing emotional variability.
• To introduce the MFE-Former model, combining self-supervised and supervised learning.
• To enhance the extraction of depression-related emotional patterns from speech.

Main Methods:

• Developed the Emotional Word Reading Experiment (EWRE) dataset.
• Generated fine-grained emotional representations using context similarity and orthogonality constraints.
• Employed a Transformer decoder to reconstruct spectral structures and decouple identity.
• Integrated multi-scale emotion change perception and Bernoulli distribution-based decision modules.

Main Results:

• MFE-Former successfully differentiates emotional expression patterns between depressed patients and healthy individuals.
• Patients with depression showed a higher tendency for negative emotional expression.
• Healthy individuals exhibited more positive emotional expression.
• Experimental results on EWRE and AVEC 2014 datasets demonstrated superior performance over state-of-the-art methods.

Conclusions:

• MFE-Former effectively detects depression from speech by analyzing emotional changes, even with varying emotional patterns.
• The method's ability to decouple speaker identity enhances the sensitivity to emotional cues.
• MFE-Former offers a promising advancement in computational approaches to mental health assessment.